Disposability

Maximize robustness with fast startup and graceful shutdown. A twelve-factor process can be started or stopped at a moment's notice. Platforms do this all the time: to deploy a new release, to scale down at night, or to move work off a failing machine.

That asks two things of every process:

• Start fast. A few seconds, ideally. Slow startup makes deploys and scaling sluggish, and leaves you short of capacity while a crashed process comes back.
• Stop gracefully. When asked to stop, finish what you're doing, then exit.

#How a stop works

Platforms stop a process in two steps. First a polite SIGTERM, then, if it hasn't exited after a grace period, a SIGKILL that can't be caught.

#Handle SIGTERM

Production servers like Gunicorn and Uvicorn already drain requests on SIGTERM. For a worker loop you wrote yourself, catch the signal and finish the current job:

import signal
import sys
 
running = True
 
def stop(signum, frame):
    global running
    running = False  # finish the current job, then leave the loop
 
signal.signal(signal.SIGTERM, stop)
 
while running:
    job = queue.get(timeout=1)
    if job:
        handle(job)
 
sys.exit(0)

$ docker stop shop-worker
# SIGTERM sent; the worker finishes its job
shop-worker
$ docker inspect -f '{{.State.ExitCode}}' shop-worker
0

#Be ready for a crash, too

Graceful shutdown is the polite case. Machines also lose power. Make jobs safe to run twice, and put a job back on the queue if a worker dies holding it, so a crash delays work instead of losing it.

In Kubernetes, the same SIGTERM and grace period apply when a pod is deleted, and terminationGracePeriodSeconds sets the timer. Delete a pod in the Kubernetes lab and watch a replacement take its place.